Your test is actually rather irrelevant, since it is comparing the results of weeding with a surface built from real survey data.
By contrast, when we build data from contours, we are already accepting the fact that at any point on our site, we may have vertical error that is as great as our contour interval. For the most part, error is within half the contour interval, but at any one point, the error can actually shoot all the way up to the entire contour interval.
So given that we are already starting at that point, reasonable weeding the data does not introduce significant additional error. Obviously, too much weeding can create results that exceed our acceptable tolerance. But the fact that the contours were created from Lidar data (which may have already been weeded) is irrelevant. We actually don't CARE how the original contours were created. No matter how they were created, we know we are creating a surface that introduces a lot of error, simply because of the fact that we are creating surfaces from contours.
All we need to concern ourselves with is that, after we build our new surface, the resulting contours should be "close enough" to the contours we started from, for whatever purposes we intend. The original source of the contours is irrelevant, except for the fact that it might help us assign a relative level of "reliability" to our starting contours.